A world where work can be delegated to AI without surrendering control

The vision is that more of what people and organizations imagine, intend, and need to accomplish becomes possible because AI, and in time superintelligence, can take part in doing the work.

That is already happening. AI has helped discover a drug that reached human clinical trials, generated viable viral genomes subsequently tested in laboratories, and independently arrived at scientific hypotheses that took researchers years to establish. AI systems have also begun producing mathematical results addressing longstanding open problems.

These are real contributions to discovery, but their proposed results remain subject to different forms of experimental validation, independent verification, and institutional acceptance. The distinction is fundamental: producing a result is not the same as establishing that it counts.

Intelligence is beginning to take on parts of work once constrained by human time, expertise, and capacity.

The question is no longer only what AI can do. It is who remains in charge of the work and its consequences.

Control is not the vision. It is the condition under which autonomy remains directed toward our intentions, accountable to our authority, and responsive to change.

People and organizations can assign work to human or AI participants while retaining the ability to direct the arrangement, understand what was done, and determine whether the outcome counts. Assigning the work is not surrendering it.

The thesis is about how authority over work is established and retained when the work itself, including the direction of people and other agents, can be delegated to AI.

Being in charge means three things, not a signature.

Coordination: Who takes the next step, what it depends on, and how people, agents, and machines remain part of one arrangement.

Visibility: What can be seen of the work while it happens, what it was grounded on, what decisions were made, and whether the acts stayed within their established boundaries.

Governance: Who may authorize, constrain, interrupt, or change the work, and who or what determines that the result counts.

Control is these three held together. It is not a kill switch.

Consider a self-driving car. It may decide how to navigate, but not where it is authorized to operate, when it must yield control, or what constitutes a completed trip. Those conditions belong to the arrangement governing the journey, not to the intelligence driving the vehicle. Human driving has always operated within external rules; autonomy makes the separation between performing the act and governing the work explicit.

The car can decide how to drive. It does not thereby acquire authority over the journey.

The same principle applies beyond driving. A model may discover a promising drug candidate without acquiring authority to administer it to patients. It may propose an experiment without permission to conduct it. It may produce a mathematical proof without becoming the authority that establishes its validity.

The point is not to prescribe every detail of how work is performed. The driving, the assay, the answer can be left to the participant, within the methods and constraints the work requires. Nor is it merely about writing things down. Records and specifications make coordination, visibility, and governance explicit, durable, and verifiable rather than dependent on the performer.

What we seek is goal-directed autonomy, delegated by us.

A participant may decide how to pursue a given outcome. That is independence in the means, not independence from the purpose, constraints, and authority governing the work.

The right to decide does not follow from who performs the task.

But the distinction runs deeper. It is not simply who exercises authority, but who may redefine authority, and within whose limits.

There are two levels of control.

Operational control: Who assigns tasks, sets priorities, coordinates participants, makes decisions, and directs execution within an established arrangement.

Constitutive control: Who establishes or changes the arrangement itself—its purpose, scope, participants, delegated authority, constraints, and conditions for changing or accepting the work.

Both can be delegated to AI.

An agent may direct a team that includes people, assign their work, set priorities, and make decisions they are expected to follow. The organizational hierarchy may place an agent above a human for a particular purpose. Reporting relationships determine who directs work within an arrangement, not where the authority to establish that arrangement originates.

An agent may also exercise constitutive control. It might reorganize a team, create new agents, redesign a workflow, reallocate responsibilities, or establish new decision procedures. It may even delegate some of those powers to another agent.

The difficulty is recursive.

An agent authorized to redesign a work system may delegate that authority to another agent. That agent may change the arrangement again, appoint new participants, or establish further delegations. Each change can alter not only who performs the work, but who has the power to determine how future work is governed.

Recursion is not itself a loss of control. The number of agents, the depth of delegation, and the presence of AI managers are not the decisive questions.

The danger is a delegation that becomes self-authorizing. An agent may acquire the effective power to enlarge its mandate, weaken its constraints, change the conditions under which its decisions are reviewed, or defeat the mechanisms by which its authority can be challenged. A principal can retain formal authority while losing the practical ability to exercise it.

Constitutive control can itself be delegated. An agent may change the arrangement, including the authority of other participants, provided those changes remain within the limits of its own delegation. Those limits must remain accountable to the principal, though they need not be exercised exclusively by the principal.

An agent may change the arrangement. It must not acquire the power to make its own authority unlimited merely by changing the arrangement.

Retained control is not established by authorizing the first delegation. It must remain effective after the last.

The unit this applies to is the work system, not the agent and not the model.

The work can be a drive, a scientific experiment, a business process, or a mission. The participant can be a person, an agent, or a machine. The progression can be a route, a protocol, a sequence of decisions, or an exchange among participants.

A model is not the work system. It is something a participant uses, or something that constitutes part of the participant.

This distinction matters because participants can change while the work and its obligations remain.

A different model may perform the same act. Multiple agents may collaborate. A human may intervene, approve, or take over. A manager agent may coordinate workers. A verifier may establish whether another agent’s result satisfies the required conditions.

The arrangement must remain coherent across all of them.

Consider a drug discovery program. One model identifies a biological target. Another proposes molecules. A simulation evaluates candidate interactions. Laboratory equipment produces observations. Scientists interpret the evidence, and authorized institutions determine whether testing may proceed.

The work is not any one model’s output. It is the progression from an objective through discovery, testing, evidence, and authorized decisions. Participants and methods may change without changing the obligations governing the work.

Perception, reasoning, and planning may be end to end. The work system need not be. Its purpose, boundaries, authority, fallback, and evidence must remain independently accountable, even where coordination or verification is delegated to an agent.

The work system is the object being designed, whether proposed, simulated, deployed, or running. Coordination, visibility, and governance are what allow that design to remain coherent as work is delegated and performed.

This principle applies independently of the capability of the performer, including, in principle, superintelligence, provided the principal retains meaningful means to direct the arrangement, observe its consequences, and enforce or refuse outcomes.

Where those means no longer exist, retained control is a claim, not a reality.

Performing an act does not, by itself, establish that the intended work has been completed.

An answer can be produced without resolving the question. A molecule can be discovered without becoming an approved medicine. A mile can be driven without completing the trip.

Execution and settlement are different.

The criteria for recognizing completion must be established independently of the participant performing the act.

They may be evaluated automatically, by a human, or by another authorized agent. Not every routine action requires human approval. Standing authority can permit work to proceed and outcomes to be recognized without interruption when the applicable conditions are satisfied.

Some consequences cannot be refused after the fact. A payment may already have been sent, a machine may already have moved, or information may already have been disclosed. Where an act is irreversible or consequential, governance must establish what must be checked before execution.

A fallback is not completion. An agent that reaches the boundary of its authority must stop, escalate, or follow an authorized alternative rather than invent permission to continue.

The participant performs the act. The work system establishes whether it was authorized, whether it satisfies the intended outcome, and what happens next.

Only then do six questions earn their place.

They specify the work system so that coordination, visibility, and governance can be established and checked.

1. What is the work? Its intended outcome and scope.
2. How does the work progress? Its steps, dependencies, decisions, and exceptions.
3. Who participates, and how are they related? The roles, responsibilities, and arrangements among people, agents, and machines.
4. What grounds the work? The evidence, information, assumptions, and context on which decisions may rely.
5. What gives the go-ahead to act? The authority, capabilities, constraints, and limits under which participants may proceed.
6. What makes the result official? The conditions, evidence, and authority required to recognize an outcome.

The six questions describe the arrangement. They are not six separate controls, nor do they map exclusively to the three dimensions.

Coordination directs the progression of work. Visibility establishes what can be observed and evidenced. Governance determines what is permitted, when intervention is necessary, and how outcomes are recognized.

All three operate across the work system.

The six questions are the apparatus, not the opening claim.

Defining a work system is necessary, but not sufficient.

A specified workflow does not govern an agent that can disregard it. Permission does not prove compliance. A record that something happened does not establish that it was authorized.

Retained control therefore has three operational requirements.

Specified control: The work, participants, progression, grounding, authority, constraints, and completion conditions are established.

Enforced control: The relevant boundaries are applied through runtime mechanisms, independent checks, authorization gates, or other controls appropriate to the consequences.

Evidenced control: It can be established what happened, under which authority, against which requirements, and with what result.

The work order expresses the progression. Rules establish constraints. A grant records delegated authority. A brief gives the participant its instructions and context. Their relationships must remain consistent and verifiable.

A grant alone is permission, not a governed work system. It must be bound to the work it authorizes and the constraints under which that work may proceed.

A hashed grant can establish the integrity of a declaration and bind it to a particular specification. It cannot, by itself, prove compliance. That requires enforcement, evidence, or both.

Nor does an agent’s explanation establish compliance. The consequential facts are what it was authorized to do, what it actually did, what effects followed, and whether the applicable conditions were satisfied.

Where authority passes between participants, its provenance must also be verifiable: what each participant was authorized to delegate, what was actually delegated, and whether the resulting authority remained within those limits.

The point is not the package. It is the separation between what an autonomous participant is capable of doing and what the work system has authorized it to do.

Delegating an act can lower the cost of doing the work and make previously impractical work possible.

But delegation shifts costs rather than simply removing them.

The cost of performing the act may fall. The cost of establishing and maintaining its boundaries may rise. And when those boundaries are absent or ineffective, the cost of being wrong may increase.

A work system must therefore be judged by the value of the work it enables relative to the cost of performing, coordinating, and governing that work, including the expected consequences of failure.

A system requiring human approval for every routine act can preserve authority while destroying autonomy’s economic advantage. A system without enforceable boundaries can reduce operating costs while transferring unacceptable risk to the principal.

Consider an AI research system capable of generating thousands of promising experiments. Producing another proposal may be inexpensive. Laboratory capacity, materials, verification, and the consequences of an unsafe experiment are not. The value is not in maximizing proposals, but in arranging the work so that worthwhile experiments can be selected, authorized, executed, and recognized efficiently.

The design problem is to determine which decisions can be delegated, which boundaries can be enforced automatically, which exceptions require intervention, and which outcomes can be settled under standing authority.

This is not a case for maximum control at any cost. It is a case for control proportionate to the scope, uncertainty, and consequences of delegated work.

The objective is not simply to make existing work cheaper. It is to make more ambitious work viable without allowing the cost or consequences of autonomy to overwhelm its benefit.

The promise of increasingly capable intelligence is not that humans will prescribe every step more efficiently. It is that intelligence can pursue outcomes, explore possibilities, and carry out work that people and organizations could not accomplish alone.

But the greater the independence of the performer, the more important the distinction between the capacity to act and the authority to govern the work.

We do not need to dictate every step an intelligent participant takes. We need to establish the work within which it is free to act, the authority under which it acts, and the conditions under which its actions count.

That arrangement must survive changes in performers, models, capabilities, and methods. It must allow authority to be delegated without allowing the performer to invent its own authority. And it must distinguish an act being performed from the work being fulfilled.

The vision is what becomes possible when intelligence can act on our behalf. The thesis is that autonomy can be delegated without surrendering authority over the work. The work system is the unit of design through which that separation is made operational.